EEG Reference Electrodes: Your Hidden Variable

Every Number in Your EEG Is a Lie (Sort Of)

Here's something that nobody tells you when you first start working with EEG: the voltages you see on screen are not real.

Not in the way you think, anyway. When you look at an EEG trace and see a waveform fluctuating at, say, 10 microvolts, you're not seeing the absolute electrical potential at that point on the scalp. That number doesn't exist. You can't measure it. Nobody can. Not with any EEG system ever built, not with a billion-dollar research rig, not with electrodes made of pure gold.

What you're seeing is a difference. The voltage at one point on the scalp minus the voltage at another point. Always. Without exception.

That second point, the one being subtracted, is the reference electrode. And the choice of where to put it quietly shapes everything about your data. The amplitudes. The waveform morphology. The topographic maps. The spectral power. The statistics. The conclusions.

Get the reference wrong and you can spend weeks chasing a "brain signal" that's actually an artifact of your recording setup. Get it right and the same data becomes clean, interpretable, and scientifically meaningful.

This guide is about that choice. It's the guide I wish someone had handed me before I spent an embarrassing amount of time confused about why the same brain activity looked completely different in two papers that were supposedly measuring the same thing.

Why EEG Can Only Measure Differences (And Always Will)

To understand why the reference matters, you need to understand a fundamental law of electrical measurement.

Voltage is not a thing. It's a relationship. When you say a AA battery is 1.5 volts, what you actually mean is that the potential difference between its positive and negative terminals is 1.5 volts. If you only touch one terminal with your voltmeter, you get nothing. You always need two points to measure a voltage.

EEG amplifiers work the same way. Every amplifier channel has two inputs: a positive input (where you connect the recording electrode on the scalp) and a negative input (where you connect the reference). The amplifier computes the difference between them, amplifies that difference by a factor of 10,000 or more, and outputs the result.

This means the output of every EEG channel is:

Signal = V(electrode) - V(reference)

That subtraction is baked into the physics of the measurement. You cannot escape it. You cannot "remove" the reference. You can only choose where to put it, and then understand how that choice affects your data.

Here's the part that trips people up. The reference electrode is sitting on (or near) the body. The body is not electrically silent. The reference picks up its own signals: brain activity if it's on the scalp, muscle activity if it's near muscles, cardiac signals if it's near the heart. Whatever the reference picks up gets subtracted from every single recording channel.

This is the crux of the entire reference problem in EEG. There is no electrically neutral spot on the human body. Every possible reference location introduces its own bias into the data. The question is never "which reference is perfect?" The question is "which reference introduces the least problematic bias for what I'm trying to measure?"

The Major Reference Schemes (And What Each One Does to Your Data)

Over the past 95 years of EEG, researchers and engineers have converged on a handful of reference schemes that each solve the problem differently. Each one has real strengths and real weaknesses, and understanding them is what separates someone who records EEG from someone who understands it.

Linked Earlobes (A1/A2)

This is one of the oldest and most common reference schemes. You clip or paste an electrode to each earlobe, wire the two together (or average them digitally), and use the combined signal as your reference.

Why it works: Earlobes contain almost no brain tissue and very little muscle. They're about as electrically quiet as any spot on the head. By linking both earlobes, you create a reference that's roughly centered between the hemispheres, reducing any left-right bias.

Where it breaks down: "Almost no brain activity" is not zero brain activity. Volume-conducted signals from the temporal lobes do reach the earlobes. And the two earlobes are never perfectly balanced. Physical differences in electrode contact, skin thickness, or even ear size create a small voltage offset between A1 and A2. If you simply wire them together (a "physical link"), current can flow between the two electrodes through the low-impedance bridge, potentially distorting temporal lobe signals.

Best for: Clinical EEG, general-purpose recordings, situations where you need a familiar standard that most people will accept.

Linked Mastoids (M1/M2)

Similar to linked earlobes, but the electrodes sit on the mastoid processes, the bony bumps behind your ears. You can physically link them or average them digitally during recording.

Why it works: Mastoids are bony, which means relatively low muscle artifact compared to other scalp locations. Like earlobes, they provide a roughly bilateral, near-midline reference. Many clinical protocols and large-scale research studies use linked mastoids, making your data compatible with a huge body of existing literature.

Where it breaks down: The mastoids are closer to the brain than the earlobes, so they pick up more volume-conducted cortical activity, especially from temporal and posterior regions. This means that temporal and parietal signals in your data will be attenuated, because the reference is subtracting some of the same activity that those channels are trying to record. For studies focused on auditory processing or temporal lobe activity, this is a real problem.

Best for: Research that needs to align with published standards, auditory ERP studies (with caveats), and recordings where you want a stable, well-documented reference.

Cz (Vertex) Reference

Cz is the electrode position at the very top of the skull, midway between the nasion and the inion and midway between the two ears. Some recording systems use Cz as the online reference.

Why it works: Cz is equidistant from all other standard electrode positions, which gives it a symmetry that's mathematically appealing. It's convenient for cap-based systems because it's just another electrode in the cap.

Where it breaks down: The vertex is not electrically neutral. It sits directly over the supplementary motor area and somatosensory cortex. Any brain activity at Cz gets subtracted from every channel. Motor-related signals, somatosensory responses, and vertex sharp waves (common during drowsiness and sleep) all contaminate the reference. You end up seeing inverted copies of Cz activity injected into channels far from the vertex.

Best for: Online recording reference when you plan to re-reference offline. Cz is a reasonable choice if you know you'll apply average reference or another scheme during analysis, because you can mathematically remove its contribution.

Nose (Tip of the Nose) Reference

An electrode placed on the tip or bridge of the nose.

Why it works: The nose is about as far from the brain as you can get while still being on the head. It picks up very little cortical signal, making it a relatively "clean" reference for brain activity. It's particularly popular in studies of EEG topography because it doesn't bias left vs. right hemisphere or anterior vs. posterior distributions.

Where it breaks down: Two words: eye blinks. The nose is extremely close to the eyes, and eye movements and blinks produce massive electrical artifacts (electrooculographic, or EOG, signals) that swamp anything the brain is doing. Every blink injects a huge transient into every channel of your recording. Nose reference also picks up some frontal brain activity through volume conduction, though less than you might expect given the distance from frontal cortex.

Best for: Studies of topographic voltage distribution, research where eye movements are tightly controlled, and situations where you need a reference that's genuinely "off the brain."

Average Reference (Common Average Reference, or CAR)

This is not a physical electrode at all. It's a mathematical operation performed on the data after recording (or sometimes in real-time on the device). You compute the average voltage across all recording channels at each moment in time, then subtract that average from every channel.

The logic: If you have enough electrodes evenly distributed across the head, the average should theoretically approximate zero potential, because positive and negative fields generated by any brain source should cancel out when summed over the whole head. Subtracting this average should remove common noise while preserving localized signals.

Why it works: Average reference eliminates any signal that's common to all electrodes, including much of the reference contamination from whatever electrode was used as the online reference. It doesn't require any additional hardware. And with high-density arrays (64 channels or more), it provides an excellent approximation of a true "zero" reference.

Where it breaks down: This is where channel count matters enormously. The mathematical proof that average reference approximates zero potential assumes that your electrodes uniformly sample the entire surface of the head. No real EEG system does this. Consumer and portable systems with 8 to 32 channels sample a fraction of the scalp. With sparse coverage, the average is biased toward the brain regions that happen to be covered, and subtracting it can distort signals rather than clean them.

Think of it this way. If all your channels are over the frontal and central cortex (as is common in many consumer devices), then your "average" is really just a "frontal-central average." Subtracting it will artificially inflate signals from posterior regions and suppress signals from frontal regions. The fewer channels you have, the worse this problem gets.

Best for: High-density research arrays with 64 or more channels and approximately uniform scalp coverage. Should be used with caution when working with low-density or asymmetrically distributed electrode arrays.

REST (Reference Electrode Standardization Technique)

This is the most recent addition to the reference toolbox. REST uses a mathematical model of the head (a boundary element model or spherical head model) to estimate what the EEG would look like if it were referenced to a point at infinity, a theoretical point so far from the brain that it has zero potential.

Why it matters: REST attempts to solve the fundamental problem that plagues all other references: the fact that no physical location on the body is electrically neutral. By computing a "point at infinity" reference, REST should theoretically give you reference-free EEG.

Where it breaks down: REST depends on the accuracy of the head model. Different head model assumptions produce different results. It's computationally heavier than simple average reference. And it's less widely adopted, so using REST can make your data harder to compare with the existing literature.

Best for: Researchers who understand forward modeling and want the most mathematically principled reference possible. Increasingly used in source localization studies.

How Reference Choice Changes What You See

This is where the rubber meets the road. Same brain, same moment, same neural activity. But switch the reference and the data looks completely different.

Let me walk you through a concrete example that illustrates why this matters.

Imagine a strong alpha rhythm (10 Hz oscillation) generated in the occipital cortex. With a nose reference, the posterior electrodes (O1, O2, Oz) show large-amplitude alpha brainwaves, and the frontal electrodes show relatively flat traces. This is what you'd expect: the source is in the back of the head, so the electrodes closest to it pick up the strongest signal.

Now re-reference the same data to Cz. Suddenly the posterior alpha looks different. Its amplitude might decrease, because Cz itself was picking up some volume-conducted alpha, and subtracting Cz partially cancels the signal at the posterior sites. The frontal channels now show an inverted alpha rhythm, not because there's an alpha source in the frontal lobe, but because Cz contained alpha that's now being subtracted from frontal electrodes that didn't contain much alpha on their own.

Same brain. Same recording. Two completely different pictures.

This is not a hypothetical scenario. This exact confusion shows up in published literature. Researchers have drawn opposite conclusions about the topographic distribution of brain signals because they used different reference schemes without fully accounting for the reference's contribution.

Re-Referencing: Fixing It in Post

Here's the good news. Because EEG is pure subtraction, you can mathematically transform your data from one reference to another after recording. This process, called re-referencing, is one of the most important preprocessing steps in EEG analysis.

The math is simple. If you recorded your data with reference R1 and you want to transform it to reference R2, the operation is:

V(channel, new ref) = V(channel, R1) - V(R2, R1)

In English: take the signal at each channel (referenced to R1) and subtract the signal at R2 (also referenced to R1). The R1 terms cancel, and you're left with data referenced to R2.

For average reference, you compute the mean across all channels at each time point and subtract it from every channel. For linked mastoids, you average M1 and M2 and subtract that average from all channels.

The Crown captures brainwave data at 256Hz across 8 channels. All processing happens on-device. Build with JavaScript or Python SDKs.

Explore the Crown

Re-Referencing in Practice

Every major EEG analysis toolkit supports re-referencing. Here's what the workflow typically looks like.

In MNE-Python:

You load your raw data, set the montage, and call set_eeg_reference() with your chosen scheme. For average reference, you pass 'average'. For linked mastoids, you pass the channel names of M1 and M2. MNE handles the math and updates the data in place.

In EEGLAB (MATLAB):

The pop_reref() function lets you re-reference to any channel or set of channels. You can also compute average reference with a single function call. EEGLAB will warn you if your channel count is low enough that average reference may be unreliable.

In BrainFlow (Python/JavaScript):

If you're streaming from a Neurosity Crown through BrainFlow, you get raw EEG data referenced to the device's onboard reference. You can apply any re-referencing scheme in your processing pipeline before analysis. BrainFlow's DataFilter class provides the signal processing tools you need.

The critical rule: You must know your original recording reference before re-referencing. If you don't know what was subtracted during recording, you can't correctly undo it. Always document the online reference used during data collection.

How Many Channels Do You Need for Average Reference?

This question comes up constantly, and the answer has real implications for consumer EEG devices.

A 2001 paper by Junghoefer and colleagues showed that average reference becomes a reasonable approximation with around 32 channels, and a good approximation with 64 or more channels that provide roughly uniform coverage. Below 32 channels, the bias introduced by incomplete spatial sampling can be worse than the bias introduced by a single physical reference electrode.

For a device like the Neurosity Crown with 8 channels, average reference requires careful thought. The channels cover specific regions (CP3, C3, F5, PO3, PO4, F6, C4, CP4), giving bilateral coverage across frontal, central, centroparietal, and parieto-occipital areas. That's good spatial diversity for 8 channels. But it's still 8 channels, not 64.

In practice, many developers working with the Crown keep the device's native reference for real-time applications like focus detection and neurofeedback, where the onboard algorithms are already optimized for the recording configuration. For offline analysis or research applications, they may re-reference to a specific scheme that aligns with their analysis approach.

The honest answer: with 8 channels, no reference scheme is perfect. What matters most is that you understand which one you're using and account for its biases in your interpretation.

Reference Considerations for BCI Development

If you're building brain-computer interfaces, reference choice affects classification accuracy in ways that are often underappreciated.

For motor imagery BCIs (detecting imagined left vs. right hand movements), the reference matters because the signals of interest originate from the motor cortex, roughly under C3 and C4. A Cz reference is problematic here because Cz sits directly between the two sources and will partially cancel them. Linked ears or mastoids work better because they're distant from the sources of interest.

For attention and focus detection, frontal asymmetry and spectral power ratios are common features. A reference on the nose or at the ears minimizes the chance that the reference itself contributes frontal activity to the signal.

For event-related potential (ERP) analysis, including P300 BCIs, the reference can shift the latency and amplitude of components. Average reference is commonly used in ERP research, but only when channel count supports it. Linked mastoids are the standard alternative.

The "I Had No Idea" Fact About Reference Electrodes

Here's something that genuinely surprised me when I first learned it.

In the early days of EEG, researchers didn't think much about the reference. They just needed a second wire. So they clipped it wherever was convenient. An ear. A chin. Sometimes the ankle. Hans Berger, who recorded the first human EEG in 1929, initially used the back of the neck as his reference.

For years, different labs used different reference locations without always reporting which one they'd used. Published papers would describe the recording electrodes in detail and then say something like "referenced to the ear" without specifying which ear, or whether the ear electrode was on the lobe or the mastoid.

This means that a chunk of early EEG literature, the foundational studies that established our understanding of alpha rhythms, sleep staging, and abnormal EEG patterns, was recorded with inconsistent and sometimes undocumented reference schemes. The data in those classic papers is a mixture of brain activity and reference contamination that can't be fully disentangled after the fact.

It wasn't until the 1950s and 1960s that standardization efforts (led partly by the same Herbert Jasper who created the 10-20 system) established conventions for reference electrode reporting. And it wasn't until the 1980s and 1990s that average reference and mathematical re-referencing became standard tools in the analyst's toolkit.

The entire field of EEG spent its first half-century not fully accounting for the thing that literally defines what the signal means. If that doesn't make you pause before casually accepting any EEG measurement at face value, nothing will.

What This Means for Your Work

Whether you're analyzing brain data from a Neurosity Crown, building a BCI application, or diving into EEG research for the first time, the reference is your invisible partner in every measurement. It shapes your data before you ever see it.

The practical takeaways are straightforward.

Know your reference. Whatever device or system you're using, find out what the recording reference is. Check the documentation. Check the hardware specifications. Don't assume.

Think about what the reference is picking up. Is your reference near the brain signals you care about? If so, those signals will be attenuated in your recording. Is the reference near sources of artifact? If so, that artifact will appear (inverted) in every channel.

Consider re-referencing. If your original reference isn't optimal for your analysis, transform the data. Average reference is powerful but needs sufficient channel coverage. Linked ears or mastoids are the safe default for most applications.

Report your reference. If you're publishing results, sharing data, or building tools that others will use, document the reference scheme. Future-you (and everyone who uses your work) will thank you.

The reference electrode doesn't get the attention it deserves. It's not as exciting as machine learning classifiers or real-time neurofeedback loops. But it's more fundamental than any of those things. Every algorithm, every feature extraction, every statistical test you run on EEG data sits on top of a reference choice. Get that foundation right, and everything built on it becomes more trustworthy.

Get it wrong, and you're building on sand.